New technologies for listening to bats

As well as being fascinating creatures with a unique sensory ecology, bats are also potentially useful indicator species whose population trends may help to provide broader information about the health of ecosystems. To mark Halloween, Ella Browning and Rory Gibb describe new research developing smart tools to more effectively and accurately monitor bat populations across the globe. Ella Browning and Rory Gibb are both PhD students working within Prof. Kate Jones’ lab at the Centre for Biodiversity and Environment Research at University College London. Ella’s PhD research focuses on developing new tools for monitoring bats using passive acoustics, while Rory works on modelling impacts of global change and biodiversity loss on zoonotic disease transmission.

Bats are fascinating and often misunderstood creatures: the only mammals to have evolved true powered flight, possessed of extraordinary sensory systems, and important providers of ecosystem services including pest control and pollination. Bats are also sensitive to human pressures, meaning that they’ve often been proposed as bioindicator species, whose populations can tell us something about how ecosystems are faring more broadly. It’s therefore vital to become more effective at surveying and monitoring bat populations, and to understand how new technologies can assist in doing so. Bats navigate and hunt using echolocation, meaning that they’re continually releasing acoustic information into the environment, in the form of regular ultrasonic calls. Although too high in frequency for humans to hear, these can be captured by specialised ultrasound detectors, and the resulting data analysed to assess trends in behaviour, activity and populations of bat species. The Bat Conservation Trust (BCT) has been running the National Bat Monitoring Programme since 1997, enlisting members of the public to assist in regular acoustic surveys with handheld bat detectors (as well as winter and summer roost counts). Over the last decade, the iBats programme founded by Prof. Kate Jones, and also the Norfolk Bat Survey, have expanded on this approach, recruiting citizen scientists to use ultrasonic sensors to record thousands of hours of survey audio data for subsequent analysis.

Lesser horseshoe bat
Lesser horseshoe bat. Photo by Jessicajil

Much of the recent research in Prof. Jones’ group at University College London has focused on improving technologies and methods for acoustic bat monitoring, often with the help of citizen scientists. This involves developing an efficient open-source pipeline for monitoring: starting with sensors in the field listening for bats, moving through audio processing, and finally analysing ecological patterns and trends. Each of these steps involves its own challenges. For example, ultrasonic detectors are often very expensive (often thousands of pounds each) which often limits the scale of monitoring programmes. Then, once the audio data are collected, the difficulty then becomes how to effectively analyse thousands of hours of audio recordings. Developing an affordable and open-source monitoring pipeline – whose underlying methods are transparent, to improve scientific understanding of their limitations – involves finding ways to address these interlinked challenges, often through collaboration between ecologists and computer scientists.

naturesmartcities
Pipeline from Nature Smart Cities
audiomoth
AudioMoth sensor

How to avoid the high cost of most full-spectrum ultrasound detectors? The recent emergence of low-cost acoustic sensors based on inexpensive components offers exciting new possibilities. Recently, our group has been working with the AudioMoth team at Oxford and Southampton universities to explore using their newly developed sensors for bat monitoring. AudioMoths use a microelectro-mechanical systems  microphone (commonly referred to as MEMS microphones), which are power savvy and sensitive, making them capable of successfully recording even the highest frequency bat calls in the UK – the lesser horseshoe bat, at over 100kHz. So far, these new sensors have been used in projects including gunshot detection in tropical forests (for monitoring poaching), surveying endangered bats in Cuba, and conducting a comprehensive bat survey of the island of Madeira. Perhaps most importantly, they’re affordable, at around £40 per unit. This creates the potential to massively increase the spatial and temporal coverage of bat surveys in the UK and globally, and also to get citizen scientists involved in data collection: both engaging people with the fascinating world of bats, and providing valuable data for understanding biodiversity trends.

Now that collecting vast amounts of audio data is more feasible, the next challenge becomes how to efficiently manage and analyse it. Before we can analyse activity and population trends, the first step involves finding all of the bat calls in each recording (detection) and identifying the species that produced them (classification). Doing this manually is massively time-consuming, so in the last few years much of our group’s work has focused on developing open-source machine learning algorithms to quickly and accurately find bat calls in audio recordings: voice recognition software for bats, if you like. For example, in a recent study led by Dr. Oisin Mac Aodha, we trained a convolutional neural network (deep learning) to automatically detect search-phase bat echolocation calls, using lots of examples of calls labelled by volunteers on our citizen science project Bat Detective. We found that using this method, and training the algorithms on noisy, real-world audio data (from the iBats project), enabled our tool to substantially outperform other softwares, especially at detecting fainter or partly-obscured calls. The tool also runs quickly, meaning that it can rapidly process large amounts of raw audio. Its improved performance on noisier data may be especially be useful for the future, as we begin to use low-cost sensors (such as AudioMoths) whose signal-to-noise ratio and frequency sensitivity may be more variable than in more expensive sensor models.

Finally are the exciting, broader projects that combine all of these approaches. The recently-launched Nature-Smart Cities project is a collaboration between a number of organisations, including UCL and Intel, to develop the first real-time sensor network for monitoring bats. The team developed a network of smart bat sensors using Intel’s Edison computer, which is now installed across the Queen Elizabeth Olympic Park in London. Each sensor applies the deep learning detection/classification tools to the ultrasonic audio-stream in real-time, and automatically reports bat detections to the cloud and an interactive website. Since the sensors were installed this summer, the system has been monitoring bats across the park. The data they collect may ultimately help to answer questions about the drivers of bat activity, and how bats are responding to human disturbances in a city environment. But importantly, it has also enabled the public to log onto the website each night, and watch bat activity unfold across the sensor network in real time – a fantastic way to access the otherwise hidden world of bats.

Leave a comment